Skip to main content

GradientAI Provider Guide

GradientAI (DigitalOcean) is a Tier-2 catalog provider: OpenAI-wire-compatible with no behavioural quirks, so its entire integration is one JSON file (src/lib/providers/catalog/gradientai.json) rather than hand-written code. That file is the source of truth for everything on this page.

Verification status: this entry is docs-verified only, not yet live-verified. The vendor's GET /v1/models needs a key (an unauthenticated GET answers HTTP 401), so the model ids come from DigitalOcean's public docs and the roster has not been checked. No account was created and no API key was used to build it. evidence.liveMatrix is null until someone runs the live capability matrix with a real key (see Verification status below).

DigitalOcean's docs now brand the product "Inference"; the API reference describes the serverless API as part of "DigitalOcean Gradient™ AI Agentic Cloud".

Key Facts​

  • Provider id: gradientai
  • Protocol: OpenAI-compatible (/chat/completions)
  • Base URL: https://inference.do-ai.run/v1
  • Default model: llama-4-maverick
  • Models in catalog: 12, copied from the models page (retrieved 2026-09-29), which states "DigitalOcean Inference supports more than 70 foundation, embeddings, and reranking models"
  • Streaming: supported — the Serverless Inference API reference documents stream: "If set to true, the model response data will be streamed to the client as it is generated using server-sent events."
  • Tool calling: model-dependent — the API reference documents tools and tool_choice request parameters for chat completions, and the models page carries a "Tool calling" usage note on individual models (for example glm-5.2, glm-5.3-flash and mimo-v2.5-pro in this catalog)
  • Tools while streaming: not declared (false) — the API reference documents stream and tools as separate request parameters
  • Structured output: not declared (false) — the models page carries a "Structured outputs" usage note on the glm-5.2, glm-5.3-flash and mimo-v2.5-pro rows
  • Structured output + tools together: not declared (false) — no combined probe was possible without credentials
  • Embeddings: not declared (false) — the endpoint list documents POST /v1/embeddings; the flag records that NeuroLink's generic catalog provider class does not implement embed() / embedMany(), not what the vendor's API offers
  • Thinking: not declared (false) — the Use Reasoning page describes a reasoning_effort request parameter for models that support reasoning, and the flag here is provider-wide
  • Billing: no-free-tier — the schema's closest member to the vendor's statement "You must maintain a positive prepaid account balance to send serverless inference requests" (pricing page)
  • Key format: none declared (apiKeyFormat: null)

Quick Start​

1. Get an API key​

  1. Visit: https://cloud.digitalocean.com/model-studio/manage-keys (the Control Panel page that https://docs.digitalocean.com/products/inference/how-to/manage-model-access-keys/ names for creating and managing model access keys; it shows the DigitalOcean login when signed out) and sign in to a DigitalOcean account
  2. Click Create model access key, enter a name, and copy the secret key when it is shown: the docs state "Copy and store it securely before closing the window."
  3. Prepaid balance: the docs state "Serverless Inference requires a positive prepaid account balance before you can send inference requests." (https://docs.digitalocean.com/products/inference/how-to/manage-serverless-inference-prepayment/)
  4. Set GRADIENTAI_API_KEY in your .env file

2. Configure​

export GRADIENTAI_API_KEY=your-model-access-key
export GRADIENTAI_MODEL=llama-4-maverick   # optional — overrides the default model
export GRADIENTAI_BASE_URL=https://inference.do-ai.run/v1   # optional — proxy or self-hosted gateway

3. Use it​

import { NeuroLink } from "@juspay/neurolink";

const neurolink = new NeuroLink();

const result = await neurolink.generate({
  input: { text: "Explain context windows in one paragraph." },
  provider: "gradientai",
  model: "llama-4-maverick",
});

console.log(result.content);
# CLI
npx @juspay/neurolink generate "Hello" --provider gradientai

Per-request credentials use the standard NeuroLink pattern:

await neurolink.generate({
  input: { text: "Hello" },
  provider: "gradientai",
  credentials: { gradientai: { apiKey: process.env.GRADIENTAI_API_KEY } },
});

Models​

ModelContextMax outputVision$/M in · out (cache read)Notes
llama-4-maverick ⭐128,00016,384—$0.25 / $0.87NeuroLink default. Models page row "Llama 4 Maverick 17B 128E Instruct"; the Chat Completions page uses model="llama-4-maverick" in its Python OpenAI SDK example.
deepseek-v4.1-flash1,048,5761,048,576yes$0.30 / $1.20 ($0.006)NeuroLink fallback and models.visionModel. Usage notes "Prompt caching", "Native vision (text and images)".
deepseek-v4-pro-08131,048,5761,048,576—$1.32 / $3.96 ($0.044)Usage notes "Input context window of up to 1M tokens", "Prompt caching".
deepseek-v4-flash-07311,048,5761,048,576—$0.14 / $0.28 ($0.028)NeuroLink fallback. Usage notes "Input context window of up to 1M tokens", "Prompt caching".
deepseek-3.2163,840163,840—$0.50 / $1.60 ($0.15)Models page row "DeepSeek V3.2". Usage note "Prompt caching".
glm-5.2262,144262,144—$1.40 / $4.40 ($0.21)NeuroLink fallback. Usage notes "Text only", "Tool calling", "Structured outputs", "Reasoning".
kimi-k2.6262,144262,144yes$0.95 / $4.00 ($0.19)NeuroLink fallback. Models by Provider table: "both with prompt caching and multimodal vision support"; Features page lists it with "Text and image inputs".
kimi-k31,048,576Not publishedyes$3.00 / $15.00 ($0.30)Usage notes "Chat Completions API for sending prompts for serverless inference", "Native vision (text, images)"; Features page lists it with "Text and image inputs".
glm-5.3-flash1,048,5761,048,576yes$0.15 / $0.50 ($0.03)Usage notes "Multimodal: text, image, and video inputs", "Tool calling", "Structured outputs", "Reasoning".
mimo-v2.5-pro262,144262,144—$0.80 / $3.00 ($0.16)Models page row "MiMo V2.5 Pro". Usage notes "Text only", "Tool calling", "Structured outputs", "Reasoning".
nemotron-3-ultra-550b131,072131,072—$0.90 / $1.70Models page row "Nemotron 3 Ultra". Usage note "Evaluations judge model".
gemma-4-31B-it256,0008,192—$0.18 / $0.50 ($0.036)NeuroLink fallback. Row "Gemma 4" on the models and pricing pages. Usage note "Prompt caching".

Context window, max output tokens and usage notes are from the models page; prices are from the pricing page (USD per 1M tokens, Standard processing mode). Both were retrieved 2026-09-29. The ids have not been compared with a live GET /v1/models roster. On the models page the glm-5.2 and mimo-v2.5-pro rows show context window 262,144 in the table and the usage note "Input context window of up to 1M tokens"; the entry records the table figure.

models.defaultContextWindow (32,768) and models.defaultMaxOutputTokens (4,096) are placeholders the vendor does not publish.

Vision: vision: true is set on the four rows above marked "yes", where the models page, the Features page or both state image input for that model. The entry sets vision: false on the other rows.

Fallback order when the default is unavailable: deepseek-v4.1-flash → deepseek-v4-flash-0731 → glm-5.2 → kimi-k2.6 → gemma-4-31B-it. The runtime fallback model name the loader derives (fallbacks[1]) is deepseek-v4-flash-0731.

Verification status​

Tier-2 onboarding requires evidence before a provider is accepted, and pnpm run verify:provider-onboarding gates it in CI. This is what the catalog currently records for GradientAI — docs-verified only, not live-verified:

ProbeResult
Rosterunauthenticated GET https://inference.do-ai.run/v1/models answered HTTP 401 on 2026-09-29 without a key, so model ids are taken from the models page (retrieved 2026-09-29); roster not verified
Auth-failure shapethe 401 body was {"id": "Unauthorized", "message": "Unable to authenticate you" }; authProbe records status 401, code Unauthorized; errorRules covers 401, 429 (body documented in the API reference) and the 404 "Model not found" response described on the Model Support Policy page
Chat routeunauthenticated GET https://inference.do-ai.run/v1/chat/completions answered HTTP 405 with an empty body on 2026-09-29; the documented verb is POST and no POST was sent
Billingprepayment page: "Serverless Inference requires a positive prepaid account balance before you can send inference requests." — 2026-09-29
Tools / structured outputtools and tool_choice are in the API reference; "Tool calling" and "Structured outputs" appear as per-model usage notes on the models page; neither was exercised live, and no combined tools+schema request was sent — structuredOutputWithTools stays false
Live capability sweepnot run. evidence.liveMatrix is null. Before treating this provider as production-ready, run npx tsx test/continuous-test-suite-provider-matrix.ts --provider=gradientai with a real key and record the result.

Do not treat this entry as equivalent to a live-verified Tier-2 provider (e.g. FriendliAI, Novita AI) until that live matrix has been run and evidence.liveMatrix is filled in.

Troubleshooting​

Each row restates what a DigitalOcean page says; the source page is named in the Symptom column.

SymptomCauseFix
HTTP 401 Unable to authenticate you (API reference)The API reference describes 401 as "Authentication failed due to invalid credentials."Authenticate with a model access key or a DigitalOcean personal access token (overview)
Key stops working (model access keys page)"Regeneration takes effect immediately, with no grace period during which both the old and new keys are valid.""update any application that uses the key as soon as you regenerate it"
Access suspended (prepayment page)"When your balance reaches $0, DigitalOcean suspends your access to Serverless Inference until you replenish it.""add a prepayment balance manually or enable auto-reload"
HTTP 404 Model not found (Model Support Policy)"The model is no longer accessible. Requests to these model IDs return a 404 Model not found error." (for deprecated models)"Update the model ID parameter in your code to the new model ID"
HTTP 429 (API reference)The API reference describes 429 as "The API rate limit has been exceeded."—

See also​